Chapter 10 Lessons Learnt from the ACA
This chapter contains some caveats and advice, summarising the lessons that we’ve learnt from the ACA.
10.1 Modelling
- Be careful with the variance/precision hyperpriors! These can have a very large impact on the degree of smoothing and ultimately how plausible the results are.
- Sensitivity analysis is important.
- Check for under-/over-smoothing and plausibility of estimates.
- Comparing to the BYM as a benchmark can be useful.
- Optimisation is important.
- If possible, use optimised software, e.g. CARBayes for ease of implementation and computational speed. But be aware of limitations (e.g. the BYM model only gives estimates of the sum, \(S+U\), not \(S\) and \(U\) separately.
- If using software like WinBUGS, optimise the code:
- Only monitor vital parameters. Parameters like the SIR and EHR can be recovered post-estimation. E.g. \(SIR_i = \exp(\beta_0 + \beta_1 x_i + S_i + U_i)\) for the BYM model (see Sections 7.3 and 8.3).
- Separate the processes of estimation and analysis. That is, run the code to estimate the posterior for all cancers in batch mode, saving MCMC output after each cancer. Then run another batch script to render summary plots. This reduces potential run-time errors from interfering with the modelling process, and also enables longer runs of the MCMC chains if it appears burn-in is too short.
10.2 Visualisation
- Plotting empirical densities for ratio-scale parameters like SIR and EHR is problematic - the area under the curve will not appear to sum to 1, and adjacent densities can appear to have vastly different areas under the curve. Essentially, densities are not designed for this scale. Potential solutions:
- Just use credible intervals to summarise the variance of the estimates;
- Plot the density on a linear scale; or
- Plot the density of the log-SIR or log-EHR and relabel the x-axis by exponentiating the values, transforming the plot to the ratio scale.
- When creating maps, plot all ratio-scale parameters on the linear scale (by taking logarithms). This ensures a consistent colour gradient between ‘break-points’.
- Choose colour palettes which are colour-blind friendly and have high-contrast (to accentuate differences).
- Consider ways to emphasise spatial patterns for small, concentrated areas (e.g. major cities) without obscuring the national picture. For static maps, this is traditionally achieved by showing map insets. For a digital product like the ACA, alternative options are possible: barcode plot, cartogram, etc.